In silico purification improves DNA methylation-based classification rates of pediatric low-grade gliomas

Abstract

DNA methylation-based classification using the Heidelberg Classifier is a state-of-the-art data-driven method for molecular diagnosis of central nervous system (CNS) tumors. However, many pediatric low-grade glioma (pLGG) samples fail to yield a confident methylation-based classification, often suspected due to low tumor cell content. Here, we present a rapid, reference-based in silico purification framework that systematically removes the epigenetic signatures of five non-malignant cell types—microglia, monocytes, neutrophils, T cells, and neurons—from tumor profiles to enable classification of previously non-classifiable pLGG samples. To validate our approach, we analyzed paired DNA methylation profiles from the same biopsy, where one was initially classifiable and the other was not. After purification, predictions for all newly classifiable samples matched the classification of their corresponding initially classifiable counterparts (9/9, 100%). Application of our method to two independent pLGG cohorts allowed confident classification in 24.1% (26/108) and 22.7% (5/22) of previously non-classifiable cases. In conclusion, our in silico purification framework enables confident classification of previously non-classifiable pLGG samples, supporting accurate molecular diagnosis and timely clinical decision-making, and can seamlessly be integrated into current classification workflows. Its independence from tumor type, classifier, and reference signatures further suggests the potential for broader application to other low-purity tumor types.